# Automate Browser Workflows With Agent Refs

Manual browser testing and research slows agents when pages require clicks, forms, or screenshots. This skill gives Claude, Codex, and Claude Code repeatable browser control through belt, element refs, snapshots, and capture tools.

## Install

```bash
npx skillstore add halt-catch-fire/agent-browser
```

## Metadata

- Status: approved
- Slug: halt-catch-fire-agent-browser
- Skillstore revision: r1
- Version status: missing
- Tree hash: 1a1be1f31412d1948c6a0cc3884f0cdf3a219df1255ea567115d33a7cee68cb4
- Author: halt-catch-fire
- GitHub username: halt-catch-fire
- License: MIT
- Repository: https://github.com/halt-catch-fire/skills/tree/main/tools/utilities/agent-browser/
- Ref: a06681402992ceae98ba04d54cfd4ab004862696
- Supported tools: Claude, Codex, Claude Code
- Audit status: complete
- Agent install advisory: confirmation\_required
- Manual install advisory: allowed
- Artifact signature: available
- Audit attestation: unavailable
- Human verification: not\_verified
- Risk factors: external\_commands, network, filesystem
- Quality score: 38
- Quality tier: warning
- Public page: https://skillstore.pages.dev/skills/halt-catch-fire-agent-browser
- Manifest: https://skillstore.pages.dev/api/skills/halt-catch-fire-agent-browser/manifest

## Capabilities

- Open browser sessions with custom viewport, user agent, proxy, and video options.
- Snapshot pages into compact @e element refs for clicks, fills, selects, drags, and uploads.
- Capture viewport screenshots, full-page screenshots, and session videos.
- Execute JavaScript in the page context for extraction and inspection.
- Provide shell templates for form automation, authenticated sessions, and content capture.
- Document proxy, authentication, 2FA, and ref lifecycle workflows.

## Use Cases

- Test Interactive Web Flows: Run login, form, and navigation flows while capturing screenshots or video evidence.
- Collect Web Page Evidence: Open pages, extract visible text and links, and save screenshots for review.
- Document Browser Workflows: Record repeatable browser actions with cursor indicators for support or training materials.

## Prompt Templates

### Open And Inspect A Page

```
Open the target page, summarize the title and visible interactive elements, then close the browser session.
```

### Fill A Simple Form

```
Use the latest snapshot to identify form fields, fill the provided values, submit the form, and report the result.
```

### Capture Research Evidence

```
Open each approved URL, capture a full-page screenshot, extract visible text and links, and list the saved artifacts.
```

### Automate An Authenticated Workflow

```
Use the provided credentials and consented target site to log in, re-snapshot after each navigation, complete the task, and avoid recording sensitive pages.
```

## Limitations

- Requires the external belt or inference.sh CLI and network access.
- Browser activity, screenshots, videos, and page data may pass through a remote service.
- Element refs become stale after navigation or DOM changes.
- Users must supply credentials, proxy details, and target URLs responsibly.

## Best Practices

- Use only approved target URLs and close every browser session after the task.
- Re-snapshot after navigation, form submission, modal changes, or dynamic page updates.
- Disable recording and screenshots when credentials or sensitive data are visible.

## Anti Patterns

- Do not use proxy rotation to evade rate limits or site access controls.
- Do not print cookies, passwords, TOTP codes, or session tokens in logs.
- Do not upload local files unless the target site and file contents are explicitly approved.

## Security Audit

- Audited at: 2026-07-06T16:24:19.941\+00:00
- Summary: Most static findings are Markdown examples, documentation links, or benign shell redirection, and those were marked as false positives. Confirmed risk remains high because the skill enables remote browser automation, screenshots, video recording, authenticated data extraction, arbitrary page JavaScript, file upload, and proxy routing. I found no prompt-injection language in the reviewed skill files.

## Stats

- Views: 0
- Downloads: 4
- Favorites: 0
- Popularity score: 0